github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-11-10 15:02:13 -06:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #16040] VRAM Detection Issue on Radeon 780M (gfx1035) - Kernel Page Fault #87898

New Issue
Open
opened 2026-05-10 06:33:45 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-05-10 06:33:45 -05:00
Owner
Copy Link
Copy Source

Originally created by @KotXL on GitHub (May 7, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/16040

What is the issue?

Diagnostic Report: AMD Radeon 780M (gfx1035) VRAM Detection Issue

System Overview

  • Hardware: GMKTec M7 Ultra (Ryzen 8845HS with Radeon 780M iGPU)
  • OS: Ubuntu 26.04 (Resolute) / Kernel 7.0.0-15-generic
  • BIOS Config: UMA Frame Buffer Size set to 8GB (Confirmed via memory mapping)
  • Software: Ollama v0.23.1

Problem Description

Despite the system correctly reporting 8GB of VRAM allocated to the iGPU, Ollama fails to offload layers to the GPU, reporting total_vram="0 B".

Key Diagnostic Findings

  1. Memory Confirmation:

    • cat /sys/class/drm/card0/device/mem_info_vram_total returns 8589934592 (8 GiB).
    • lspci confirms the VGA controller is active and using the amdgpu driver.

  2. Kernel Error (dmesg):
    The following error occurs when Ollama attempts to initialize the ROCm interface:

    amdgpu 0000:e5:00.0: [gfxhub] page fault at address 0x0000... from client 0x1b (UTCL2)

    This suggests a memory access violation or an incompatibility between the current ROCm runtime and the Phoenix architecture (gfx1035) in this kernel version.

  3. Ollama Log Error:

    error="runner crashed" detail="error: Cannot read /usr/local/lib/ollama/rocm/rocblas/library/TensileLibrary.dat: Illegal seek for GPU arch : gfx1035"

    The Illegal seek on the Tensile Library indicates that the library provided by the Ollama binary might not correctly support the specific sub-variant of the gfx1035 architecture without manual compilation or specific versioning.

Steps Taken (Mitigation)

  • Forced environment variable HSA_OVERRIDE_GFX_VERSION=11.0.0 (and 11.0.2).
  • Added ollama user to video and render groups.
  • Performed full apt update && apt upgrade to latest available kernel (7.0.0-15).
  • Re-installed Ollama using official install script.
  • Verified physical hardware via reboot.

Recommendations for Experts

  • Investigate if the gfx1035 target requires a specific build of ROCm libraries not present in the pre-compiled Ollama binary.
  • Check for firmware mismatches in the amdgpu stack for Ryzen 8000 series chips on Linux Server.
  • Consider building Ollama from source linking against system-native ROCm components if available.

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @KotXL on GitHub (May 7, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/16040 ### What is the issue? # Diagnostic Report: AMD Radeon 780M (gfx1035) VRAM Detection Issue ## System Overview - **Hardware**: GMKTec M7 Ultra (Ryzen 8845HS with Radeon 780M iGPU) - **OS**: Ubuntu 26.04 (Resolute) / Kernel 7.0.0-15-generic - **BIOS Config**: UMA Frame Buffer Size set to **8GB** (Confirmed via memory mapping) - **Software**: Ollama v0.23.1 ## Problem Description Despite the system correctly reporting 8GB of VRAM allocated to the iGPU, Ollama fails to offload layers to the GPU, reporting `total_vram="0 B"`. ### Key Diagnostic Findings 1. **Memory Confirmation**: - `cat /sys/class/drm/card0/device/mem_info_vram_total` returns `8589934592` (8 GiB). - `lspci` confirms the VGA controller is active and using the `amdgpu` driver. 2. **Kernel Error (dmesg)**: The following error occurs when Ollama attempts to initialize the ROCm interface: ``` amdgpu 0000:e5:00.0: [gfxhub] page fault at address 0x0000... from client 0x1b (UTCL2) ``` This suggests a memory access violation or an incompatibility between the current ROCm runtime and the Phoenix architecture (gfx1035) in this kernel version. 3. **Ollama Log Error**: ``` error="runner crashed" detail="error: Cannot read /usr/local/lib/ollama/rocm/rocblas/library/TensileLibrary.dat: Illegal seek for GPU arch : gfx1035" ``` The `Illegal seek` on the Tensile Library indicates that the library provided by the Ollama binary might not correctly support the specific sub-variant of the gfx1035 architecture without manual compilation or specific versioning. ## Steps Taken (Mitigation) - Forced environment variable `HSA_OVERRIDE_GFX_VERSION=11.0.0` (and `11.0.2`). - Added `ollama` user to `video` and `render` groups. - Performed full `apt update && apt upgrade` to latest available kernel (7.0.0-15). - Re-installed Ollama using official install script. - Verified physical hardware via reboot. ## Recommendations for Experts - Investigate if the **gfx1035** target requires a specific build of ROCm libraries not present in the pre-compiled Ollama binary. - Check for firmware mismatches in the `amdgpu` stack for Ryzen 8000 series chips on Linux Server. - Consider building Ollama from source linking against system-native ROCm components if available. ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-05-10 06:33:45 -05:00
GiteaMirror commented 2026-05-10 06:33:48 -05:00
Author
Owner
Copy Link
Copy Source

@chejh-amd commented on GitHub (May 8, 2026):

Quick note: HSA_OVERRIDE_GFX_VERSION=11.0.0 targets RDNA 3 (gfx11xx), but your 780M is gfx1035 which is RDNA 2, so that override won't help here.

The more practical path would be skipping ROCm entirely and trying OLLAMA_VULKAN=1. Vulkan doesn't depend on Tensile/rocBLAS and tends to work on a wider range of AMD GPUs including older RDNA 2 iGPUs. Worth a quick test to see if layers actually offload.

<!-- gh-comment-id:4403853647 --> @chejh-amd commented on GitHub (May 8, 2026): Quick note: `HSA_OVERRIDE_GFX_VERSION=11.0.0` targets RDNA 3 (gfx11xx), but your 780M is gfx1035 which is RDNA 2, so that override won't help here. The more practical path would be skipping ROCm entirely and trying `OLLAMA_VULKAN=1`. Vulkan doesn't depend on Tensile/rocBLAS and tends to work on a wider range of AMD GPUs including older RDNA 2 iGPUs. Worth a quick test to see if layers actually offload.
GiteaMirror commented 2026-05-10 06:33:48 -05:00
Author
Owner
Copy Link
Copy Source

@princekin0820-art commented on GitHub (May 9, 2026):

780M should be GFX1103, GFX1035 is 680M but both are not supported by official ROCM. Try ollama-for-amd before, but Vulkan perform similar but much stable than ROCm/zluda

<!-- gh-comment-id:4411372299 --> @princekin0820-art commented on GitHub (May 9, 2026): 780M should be GFX1103, GFX1035 is 680M but both are not supported by official ROCM. Try ollama-for-amd before, but Vulkan perform similar but much stable than ROCm/zluda
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No labels bug
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
GiteaMirror ninjasurge (Kaehlen Rog)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#87898